Fixed unified FakeConverts duplication (#2877)

### Changes - Fixed the `MinMax` algorithm for the `FakeConvert` insertion case. ### Reason for changes - Incorrect `MinMax` behaviour introduced in #2622. ### Related tickets - 149128 ### Tests - Manual Incorrect behaviour - two identical layers were inserted: ![image](https://github.com/user-attachments/assets/28fb5419-c9ac-400d-a8cc-bcf1d0e32241) Corrected behaviour - only one layer for each branch was inserted: ![image](https://github.com/user-attachments/assets/d0ab38e5-320a-49b3-8c02-5669b42257f4)
openvinotoolkit · Aug 9, 2024 · 8634037 · 8634037
1 parent 01b748f
commit 8634037
Show file tree

Hide file tree

Showing 5 changed files with 169 additions and 0 deletions.
diff --git a/nncf/quantization/algorithms/min_max/algorithm.py b/nncf/quantization/algorithms/min_max/algorithm.py
@@ -926,6 +926,7 @@ def filter_func(point: StatisticPoint) -> bool:
                     transformation_layout.register(
                         self._backend_entity.create_convert_insertion_command(quantization_target_point, parameters)
                     )
+                    unified_ops_list.add(quantization_target_point)
                 continue
             parameters = calculate_quantizer_parameters(unified_values, qconfig, q_group, narrow_range)
             commands = self._backend_entity.create_unified_scales_quantizers_insertion_commands(

diff --git a/tests/openvino/native/data/2024.3/reference_scales/UnifiedScalesModel_fp8_e4m3.json b/tests/openvino/native/data/2024.3/reference_scales/UnifiedScalesModel_fp8_e4m3.json
@@ -0,0 +1,58 @@
+{
+    "Conv/fc_weights_1": {
+        "scale": [
+            [
+                [
+                    [
+                        550.8624267578125
+                    ]
+                ]
+            ],
+            [
+                [
+                    [
+                        490.8214416503906
+                    ]
+                ]
+            ],
+            [
+                [
+                    [
+                        479.1072692871094
+                    ]
+                ]
+            ]
+        ],
+        "shift": [
+            [
+                [
+                    [
+                        0.0
+                    ]
+                ]
+            ],
+            [
+                [
+                    [
+                        0.0
+                    ]
+                ]
+            ],
+            [
+                [
+                    [
+                        0.0
+                    ]
+                ]
+            ]
+        ]
+    },
+    "Cos/fc_output_0": {
+        "scale": 224.00006103515625,
+        "shift": 0.0
+    },
+    "Sin/fc_output_0": {
+        "scale": 224.00006103515625,
+        "shift": 0.0
+    }
+}
diff --git a/tests/openvino/native/data/2024.3/reference_scales/UnifiedScalesModel_fp8_e5m2.json b/tests/openvino/native/data/2024.3/reference_scales/UnifiedScalesModel_fp8_e5m2.json
@@ -0,0 +1,58 @@
+{
+    "Conv/fc_weights_1": {
+        "scale": [
+            [
+                [
+                    [
+                        70510.390625
+                    ]
+                ]
+            ],
+            [
+                [
+                    [
+                        62825.14453125
+                    ]
+                ]
+            ],
+            [
+                [
+                    [
+                        61325.73046875
+                    ]
+                ]
+            ]
+        ],
+        "shift": [
+            [
+                [
+                    [
+                        0.0
+                    ]
+                ]
+            ],
+            [
+                [
+                    [
+                        0.0
+                    ]
+                ]
+            ],
+            [
+                [
+                    [
+                        0.0
+                    ]
+                ]
+            ]
+        ]
+    },
+    "Cos/fc_output_0": {
+        "scale": 28672.0078125,
+        "shift": 0.0
+    },
+    "Sin/fc_output_0": {
+        "scale": 28672.0078125,
+        "shift": 0.0
+    }
+}
diff --git a/tests/openvino/native/models.py b/tests/openvino/native/models.py
@@ -1139,3 +1139,20 @@ def _create_ov_model(self):
         result.get_output_tensor(0).set_names(set(["Result"]))
         model = ov.Model([result], [input])
         return model
+
+
+class UnifiedScalesModel(OVReferenceModel):
+    def _create_ov_model(self):
+        input = opset.parameter([1, 3, 4, 2], name="Input")
+        multiply = opset.multiply(input, self._rng.random((1, 2)).astype(np.float32), name="Mul")
+        sin = opset.sin(multiply, name="Sin")
+        cos = opset.cos(multiply, name="Cos")
+        concat = opset.concat([sin, cos], axis=0)
+        kernel = self._rng.random((3, 3, 1, 1)).astype(np.float32)
+        strides = [1, 1]
+        pads = [0, 0]
+        dilations = [1, 1]
+        conv = opset.convolution(concat, kernel, strides, pads, pads, dilations, name="Conv")
+        result = opset.result(conv, name="Result")
+        model = ov.Model([result], [input])
+        return model
diff --git a/tests/openvino/native/quantization/test_fq_params_calculation.py b/tests/openvino/native/quantization/test_fq_params_calculation.py
@@ -18,6 +18,7 @@
 from nncf.common.quantization.structs import QuantizationPreset
 from nncf.openvino.graph.nncf_graph_builder import GraphConverter
 from nncf.openvino.statistics.aggregator import OVStatisticsAggregator
+from nncf.parameters import QuantizationMode
 from nncf.quantization.advanced_parameters import OverflowFix
 from nncf.quantization.algorithms.min_max.algorithm import MinMaxQuantization
 from tests.openvino.native.common import convert_torch_model
@@ -28,6 +29,7 @@
 from tests.openvino.native.models import FPModel
 from tests.openvino.native.models import LinearModel
 from tests.openvino.native.models import MatMul2DModel
+from tests.openvino.native.models import UnifiedScalesModel
 from tests.openvino.native.models import WeightsModel
 from tests.openvino.native.models import get_torch_model_info
 from tests.shared.helpers import compare_stats
@@ -51,6 +53,14 @@ def get_fq_nodes_stats_algo(model):
                 "output_low": output_low,
                 "output_high": output_high,
             }
+        elif op.get_type_name() == "FakeConvert":
+            scale = op.input_value(1).get_node().data
+            shift = op.input_value(2).get_node().data
+
+            nodes[op.get_friendly_name()] = {
+                "scale": scale,
+                "shift": shift,
+            }
     return nodes
 
 
@@ -94,6 +104,31 @@ def test_synthetic_models_fq_scales(model_creator_func, preset, inplace_statisti
     compare_stats(ref_nodes, nodes)
 
 
+@pytest.mark.parametrize(
+    "mode",
+    [QuantizationMode.FP8_E4M3, QuantizationMode.FP8_E5M2],
+    ids=[QuantizationMode.FP8_E4M3.value, QuantizationMode.FP8_E5M2.value],
+)
+@pytest.mark.parametrize("model_creator_func", [UnifiedScalesModel])
+def test_synthetic_models_fc_scales(model_creator_func, mode):
+    model = model_creator_func()
+    quantized_model = quantize_model(model.ov_model, {"mode": mode})
+    real_nodes = [op for op in quantized_model.get_ops() if op.get_type_name() == "FakeConvert"]
+
+    ref_stats_name = model.ref_graph_name.split(".")[0] + f"_{mode.value}.json"
+    ref_stats_path = get_actual_reference_for_current_openvino(REFERENCE_SCALES_DIR / ref_stats_name)
+    ref_nodes = load_json(ref_stats_path)
+
+    assert len(ref_nodes) == len(real_nodes), "The number of the real FakeConvert nodes is not correct"
+    stat_nodes = get_fq_nodes_stats_algo(quantized_model)
+
+    # Uncomment lines below to generate reference for new models.
+    # from tests.shared.helpers import dump_to_json
+    # dump_to_json(ref_stats_path, nodes)
+
+    compare_stats(ref_nodes, stat_nodes)
+
+
 @pytest.mark.parametrize(
     "overflow_fix",
     [OverflowFix.DISABLE, OverflowFix.ENABLE, OverflowFix.FIRST_LAYER],